How Patentis scores.
Every finding is independently retrieved by two language models. Where they agree, the result enters the report with a high quorum coefficient. Where they disagree, we surface the disagreement — we never silently break ties.
Quorum routing
Claude Opus and Gemini Pro run in parallel on each claim element. A finding only ships when both models cite the same prior-art paragraph or when the disagreement is explicitly labeled in the report export.
Thresholds are tenant-configurable. Forensic mode adds source URLs, paragraph citations, and per-vendor attribution in every DOCX export.
Citations & auditability
Each cited reference includes its exact paragraph or column-line citation, exportable to Word with hyperlinks intact. We benchmark retrieval against the EPO Boards of Appeal corpus quarterly; benchmark commits are timestamped in the public repo.
Transparency commitments
- Routing logic and evaluation harnesses are published — no proprietary black box scoring.
- Subprocessors are listed by name with EU residency status on /subprocessors and updated within five business days of any change.
- We do not hide which model wrote which passage when forensic mode is enabled.